Three-lake Swedish observed-pollen transfer¶
Our research account at reports/science/sweden-lake-transfer.md explains
the decision, scores and limits. This governed page retains the
source-bound predictions, tables and reproduction commands.
The locked protocol selected Gilltjärnen, Lake Avegöl and Flinkasjön from Neotoma/EPD using site, collection-unit, dataset and age metadata before target counts were read. Each is one lake site with one source pollen dataset and one core. The source-default calibrated-age rule selects 268 of 273 samples; the five other source samples remain outside this study. The selected samples span 63 site by 500 BP bins.
The input manifest binds the Neotoma relational snapshot, the authors' 10,788-row taxonomy lookup, the 31-target axis and the previously retained 81-site DK/FI/NO training profiles. Training excludes every Swedish target site. The predictor read only training and target metadata; the separately retained sealed outcomes were opened only after the predictions and their freeze receipt had been committed at 135250049.
The producer at prediction freeze is
retained byte for byte from that commit, SHA-256
3c4321f729c5ea4cc53172b3ebf0a22b782f0e873ae88ad67225b43a0ad01b62.
Later style, typing and CLI-documentation edits changed the live producer's
file hash. The verifier authenticates the frozen prediction against the
retained producer bytes. We separately ran the current producer's prepare,
predict and score commands on the same source inputs: its input manifest,
three prepared files, prediction file and five score products all matched
their governed counterparts byte for byte. The prediction SHA-256 remained
eb10b5e7225cef3c51b544c1a25cebc264d9b5ab10477d9c4216b83a38bad909.
| Target lake | Selected samples | Scored bins | Native pollen NISP | Assigned to 31 targets | Author excluded | Unmapped |
|---|---|---|---|---|---|---|
| Gilltjärnen | 100 | 23 | 59,073 | 57,697 | 1,374 | 2 |
| Lake Avegöl | 62 | 15 | 67,949 | 67,178 | 740 | 31 |
| Flinkasjön | 106 | 25 | 111,775 | 106,240 | 1,711 | 3,824 |
| Total | 268 | 63 | 238,797 | 231,115 | 3,825 | 3,857 |
All 63 bins received the three fixed baseline predictions and have positive assigned pollen counts, yielding 189 matched score rows and zero training-support refusals. The coverage, site scores, all bin scores and taxon diagnostics retain their denominators.
| Fixed method | Equal-site mean absolute fraction error | Equal-site Hellinger distance |
|---|---|---|
| Inverse distance squared | 0.017334 | 0.286757 |
| Nearest site | 0.017932 | 0.288872 |
| Training mean | 0.020682 | 0.333694 |
Inverse distance squared has the lowest three-site mean error in this declared comparison, but nearest site has the lower error at Lake Avegöl. Only three lake groups support the equal-site aggregate, and prior Nordic outcomes informed the fixed study design. This is development evidence for direct observed-pollen subcomposition, not a confirmatory estimate, LRA reconstruction, full-pollen or vegetation prediction, or a country-wide Swedish skill claim. Basin geometry, PPE, independent vegetation truth and unresolved source periods remain separate admission questions.
Reproduce¶
From the repository root, with the retained external author-reference capsule:
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m \
bijux_pollenomics_dev.source_completion.swedish_lake_transfer prepare \
--repository . --author-source-root /path/to/landclim/author-reference \
--output artifacts/jobs/swedish-lake-transfer/replay-prepared
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m \
bijux_pollenomics_dev.source_completion.swedish_lake_transfer predict \
--prepared artifacts/jobs/swedish-lake-transfer/replay-prepared \
--output artifacts/jobs/swedish-lake-transfer/replay-predictions
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m \
bijux_pollenomics_dev.source_completion.swedish_lake_transfer score \
--prepared artifacts/jobs/swedish-lake-transfer/replay-prepared \
--predictions docs/report/regions/nordic/swedish-lake-transfer/predictions \
--output artifacts/jobs/swedish-lake-transfer/replay-scored
PYTHONDONTWRITEBYTECODE=1 .venv/bin/python \
docs/report/regions/nordic/swedish-lake-transfer/verify.py
The score command must use the committed prediction file. The report manifest binds retained evidence bytes, and verify.py checks the freeze, current Neotoma relational source identity, source denominator, score coverage and site-macro arithmetic. If that published source changes, this comparison requires a reviewed refresh before it can pass. The private author lookup is authenticated by digest and is not copied here.